Missing data should be handled differently for prediction than for description or causal explanation.

Area of Science:

• Epidemiology
• Statistics
• Data Science

Background:

• Traditional methods for missing data focus on prospective data for description or causal inference.
• Routinely collected data are increasingly used for predictive modeling, where missingness patterns can be informative.

Purpose of the Study:

• To explore and develop methods for handling missing data in predictive modeling, especially with routinely collected data.
• To address the discrepancy between missing data methods used in model development and deployment.
• To evaluate the impact of differing missing data handling on prediction model performance.

Main Methods:

• Review of existing theoretical developments in missing data handling.
• Exploration of pragmatic approaches for optimizing prediction using routinely collected data.
• Emphasis on methods that incorporate information from missingness patterns.

Main Results:

• Standard methods for missing data may not be optimal for prediction tasks.
• Discrepancies in missing data handling between model development and deployment can lead to overoptimistic performance assessments.
• Methods leveraging missingness patterns show promise for prediction.

Conclusions:

• Prediction models using routinely collected data may benefit from specialized missing data methods.
• Explicit evaluation of differing methods in development versus deployment is crucial.
• The choice of method involves a trade-off between causal principles and maximizing information use.